Dot products II

Lecture 17

Minjae Park

Auburn University
MATH 2660 - Spring 2026

February 18, 2026

Attendance

Scan the QR code or go to join.iclicker.com/MBNJ.

An inner product \(\left\langle \vec{u},\vec{v} \right\rangle\) (satisfying certain axioms) assigns a scalar to each pair of vectors \(\vec{u}\) and \(\vec{v}\).
This scalar measures how “similar” or “correlated” the two vectors are.
When we take the inner product of a vector with itself, \(\left\langle \vec{u},\vec{u} \right\rangle\) measures the “size” of the vector.
The norm of \(\vec{u}\) is defined by \(\left\lVert \vec{u} \right\rVert = \sqrt{\left\langle \vec{u},\vec{u} \right\rangle}\).

The Euclidean space \(\mathbb{R}^n\) with the dot product is the most important example of an inner product space.
In this setting, “correlation” and “norm” have geometric meanings: angles and lengths.
\(\left\lVert \vec{u} \right\rVert^2 = \vec{u} \!\cdot\!\vec{u}\) (length from the dot product).
\(\displaystyle \cos\theta = \frac{\vec{u} \!\cdot\!\vec{v}}{\left\lVert \vec{u} \right\rVert\,\left\lVert \vec{v} \right\rVert}\) (angle from the dot product).

The Gram–Schmidt process is an algorithm that constructs an orthogonal (or orthonormal) basis for the vector space spanned by a linearly independent set \(S\).
The idea: add vectors one at a time, making each new vector orthogonal to the previous ones.
We do this by subtracting all components parallel to the earlier vectors, obtained using orthogonal projections.

Let \(S=\{\vec{v}_1,\dots,\vec{v}_k\}\) be linearly independent and let \(W=\mathop{\mathrm{span}}(S)\).
Set \(\vec{w}_1=\vec{v}_1\).
For \(j\ge2\), remove all components parallel to the previous vectors: \[ \vec{w}_j = \vec{v}_j - \sum_{i=1}^{j-1} \operatorname{proj}_{\vec{w}_i}\vec{v}_j. \]
Then \(\{\vec{w}_1,\dots,\vec{w}_k\}\) is an orthogonal basis of \(W\).
Normalize each \(\vec{w}_j\) if an orthonormal basis is desired.
Visualization :: Gram–Schmidt process in 3D

In statistics, covariance (a measure of how two random variables move together) is closely related to an inner product.
Suppose we want to model a quantity that depends linearly on several variables.
If the variables are correlated, it becomes difficult to isolate the effect of a single variable.
Because correlated variables tend to move together, it is hard to determine which variable is responsible for changes in the response.
The Gram–Schmidt process removes the shared movement by producing orthogonal (uncorrelated) directions.
This allows us to measure the unique contribution of each variable.
The following article illustrates this idea using automobile MPG data:
Decorrelating Features using the Gram–Schmidt Process

Let \(\vec{u} = \langle u_1,\dots,u_n \rangle\) and \(\vec{v} = \langle v_1,\dots,v_n \rangle\).
Recall that \(\vec{u} \!\cdot\!\vec{v} = u_1v_1 + \cdots + u_nv_n\).
This equals the matrix product \[ [\,u_1\ \cdots\ u_n\,] \begin{bmatrix} v_1\\ \vdots\\ v_n \end{bmatrix}. \]
The row matrix comes from writing a column vector as a row using the transpose.

We use the transpose to convert column vectors into row vectors: \[ \vec{u}^T = \begin{bmatrix} u_1\\ \vdots\\ u_n \end{bmatrix}^T = [\,u_1\ \cdots\ u_n\,]. \]
In general, the transpose of an \(n\times k\) matrix \(A\) is a \(k\times n\) matrix \(A^T\).
If \(A=[\vec{u}_1\ \cdots\ \vec{u}_k]\), then \(A^T=\begin{bmatrix}\vec{u}_1^T\\ \vdots\\ \vec{u}_k^T\end{bmatrix}\) in block notation.
Entrywise definition: if \(A_{ij}\) is the entry in row \(i\), column \(j\), then
\((A^T)_{ij} = A_{ji}\) for \(1\le i \le k\) and \(1\le j\le n\).

Consider the \(3\times5\) matrix \[ A = \begin{bmatrix} 2 & -1 & 0 & 4 & 3 \\ 5 & 2 & -2 & 1 & 6 \\ -3 & 4 & 1 & 0 & -5 \end{bmatrix}. \]
Find \(A^T\).
What is \((A^T)_{21}\)?

Scan the QR code or go to join.iclicker.com/MBNJ.

This viewpoint gives another way to understand matrix multiplication.
When multiplying a matrix by a column vector, think of the matrix as a stack of row vectors.
Let \(A^T=\begin{bmatrix}\vec{u}_1^T\\ \vdots\\ \vec{u}_m^T\end{bmatrix}\) where \(\vec{u}_i\in\mathbb{R}^k\), and let \(\vec{v}\in\mathbb{R}^k\).
Then \[ A^T\vec{v} = \begin{bmatrix} \vec{u}_1\!\cdot\!\vec{v}\\ \vdots\\ \vec{u}_m\!\cdot\!\vec{v} \end{bmatrix}. \]

Let \[ A= \begin{bmatrix} 2 & -1 & 3\\ 0 & 4 & 1\\ -2 & 5 & 0\\ 1 & -3 & 2 \end{bmatrix}. \]
Let \[ \vec{v}= \langle 1, -2, 3, 0 \rangle\in\mathbb{R}^4. \]
Compute \(A^T\vec{v}\) using dot products.

Let \(\vec{u}_1,\dots,\vec{u}_n\) and \(\vec{v}_1,\dots,\vec{v}_m\) be vectors in \(\mathbb{R}^k\).
Let \(A=[\vec{u}_1\ \cdots\ \vec{u}_n]\) be a \(k\times n\) matrix and \(B=[\vec{v}_1\ \cdots\ \vec{v}_m]\) a \(k\times m\) matrix.
Then \(A^T\) is an \(n\times k\) matrix.
The \((i,j)\) entry of \(A^T B\) is \[ (A^T B)_{ij} = \vec{u}_i \!\cdot\!\vec{v}_j. \] for \(1\le i\le n\) and \(1\le j\le m\).
Thus matrix multiplication encodes all pairwise dot products between the columns of \(A\) and \(B\).

Let \[ A = \begin{bmatrix} 1 & 4 \\ -2 & 0 \\ 3 & -1 \end{bmatrix},\qquad B= \begin{bmatrix} 2 & 1\\ -1 & 3\\ 0 & 4 \end{bmatrix}. \]
Compute \(A^T B\) using dot products.

Let \(S=\{\vec{u}_1,\dots,\vec{u}_n\}\) be an orthonormal basis of \(\mathbb{R}^n\).
That is, \[ \vec{u}_i \!\cdot\!\vec{u}_j = \begin{cases} 1 & \text{if } i=j,\\ 0 & \text{if } i\ne j. \end{cases} \]
Let \(P=[\vec{u}_1\ \cdots\ \vec{u}_n]\) be the \(n\times n\) matrix whose columns are these basis vectors.
Then \(P^T P = I_n\).
Conversely, if an \(n\times n\) matrix \(P\) satisfies \(P^T P = I_n\), then \(P\) is called an orthogonal matrix.
Thus, an orthogonal matrix is precisely a matrix whose columns form an orthonormal basis of \(\mathbb{R}^n\).

The identity matrix \(I_n\) encodes the standard basis, which is orthonormal.
Consider the matrix \[ P= \frac{1}{3} \begin{bmatrix} 2 & -2 & 1\\ 1 & 2 & -2\\ 2 & 1 & 2 \end{bmatrix}. \]
Let its columns be \(\vec{u}_1,\vec{u}_2,\vec{u}_3\).
Check that \(P^T P = I_3\), hence \(P\) is an orthogonal matrix.
The set \(S=\{\vec{u}_1,\vec{u}_2,\vec{u}_3\}\) forms an orthonormal basis of \(\mathbb{R}^3\).